Run your online shop from ChatGPT or Claude: what it can and can’t do
ChatGPT and Claude can both work your shop’s admin through MCP. We’d let one draft products but never refund an order. Sources checked 15 September 2026.
If your shop platform publishes a connector, you can run much of your online shop’s admin from ChatGPT or Claude, and this page is about where that should stop. It’s written for UK shop owners and assumes no code, and we checked what it says about MCP and the two assistants against modelcontextprotocol.io, developers.openai.com and support.claude.com on 15 September 2026. We build Orbit Commerce, which publishes a connector of its own, so read the Orbit parts as a vendor talking.
In short
- MCP (the Model Context Protocol) is “an open-source standard for connecting AI applications to external systems”, and modelcontextprotocol.io lists Claude and ChatGPT among the applications that support it.
- Anthropic lists custom connectors on Claude’s Free, Pro, Max, Team and Enterprise plans, free users limited to one; OpenAI documents adding a custom MCP server in ChatGPT’s developer mode.
- Orbit splits assistant permissions into read, safe write and live write, and an assistant gets nothing you did not tick.
- Product descriptions, order notes and customer messages can all steer a connected assistant, Orbit’s dashboard warns, so give change permissions only to assistants you trust.
- In our worked example, 40 supplier products take 4 hours by hand (£72 at an assumed £18 an hour) against about 25 minutes checking assistant drafts (£7.50).
What do people mean by “AI running my shop”?
Usually one of two quite different products: a chatbot that talks to your customers, or an assistant that works your admin. The chatbot answers “where is my order?” on your storefront and passes you the hard ones, and when it slips, it says something wrong to a stranger. Only you see the second one. You open ChatGPT or Claude, type “which products sold best last month?”, and it reads your shop and answers. Type “put the slow ones in a 15% sale” and, if you gave it permission, it does that too. This page is about the second kind.
What is MCP, and why should a shop owner care?
MCP (the Model Context Protocol) is, in its own words, “an open-source standard for connecting AI applications to external systems” (modelcontextprotocol.io, checked 15 September 2026). That’s all it is: an agreed way for an assistant to plug into something else, the way USB-C is an agreed way to plug in a charger. The same page lists Claude and ChatGPT among the applications that support it.
For a shop owner, the standard is the useful part. If your platform publishes one connection, any assistant that speaks MCP can use it, including the one you already pay for.
What can an AI assistant do in your shop: read, or write?
Both, but only what you tick. In Orbit the permissions come in three tiers. Read lets it look and report. Safe write covers recoverable changes, like a draft product or a description. Live write is anything a customer will feel, like a price change or a refund. The tiers exist so you can say yes to the first two and no to the third. Here’s how we’d treat some everyday requests:
| What you ask for | Read or write | How we’d handle it |
|---|---|---|
| “Which products sold best last month?” | Read | Safe to leave on permanently. Nothing changes. |
| “Add these 40 products from a spreadsheet” | Write | Let it create them as drafts, read them yourself, then publish. |
| “Change the price of everything in the sale collection” | Write, live | Ask for the list of what it will change first. Approve, then run. |
| “Refund order 1042” | Write, live and irreversible | We wouldn’t grant this. Money leaving is a human decision. |
| “Rewrite every product description in a friendlier tone” | Write, bulk | Do ten, read them, then do the rest. Keep the old copy exported. |
What should you never let an assistant do without approving it first?
Anything that moves money or touches a customer’s personal data, and any change that hits hundreds of records in one go. In practice, we’d keep a hand on these five:
- Refunds and payouts. An assistant misreading “refund the damaged one” as “refund them all” is a bank transfer, not a typo.
- Cancelling or editing paid orders. Same reason, plus a customer at the other end.
- Bulk price changes without a preview. Ask what it intends to change and read the list. Every serious bulk edit should be a two-step with you in the middle.
- Deleting anything. Archive instead.
- Emailing your customers. Let it draft, and press send yourself.
Everything else is fair game, and that covers most of a normal week: reports, stock questions, product creation, descriptions, collections, redirects and feed clean-ups.
How do you connect an assistant to an Orbit shop?
You paste one address into the assistant and tick what it may do. Orbit hosts the connector, so there’s nothing to install and no configuration file to edit, and the whole thing takes about two minutes.
- In the Orbit dashboard, go to Settings → AI Assistants.
- Copy the address shown there and paste it into your assistant’s Connectors setting (some apps call it Integrations or Apps).
- The assistant sends you to an Orbit approval screen. Tick the permissions it may have: reading products, orders and customers, and making the changes you allow. It gets nothing you did not tick.
- Use it. You can edit or revoke the connection at any time from the dashboard, and changes apply within seconds.
Behind that address sit hundreds of actions across products, orders, customers, inventory, discounts, returns, shipping, pages and reports, limited to the permissions you ticked. There’s more on our AI page, and the data rules are in the AI assistant data terms (version 2026-08-31).
Does it work with ChatGPT, Claude, or both?
Both, today. OpenAI’s developer documentation describes connecting a custom MCP server to ChatGPT: turn on developer mode under Settings, then add your server URL (developers.openai.com, checked 15 September 2026). Anthropic’s support article says custom connectors using remote MCP “are available on Claude, Cowork, and Claude Desktop for users on Free, Pro, Max, Team, and Enterprise plans”; free users are limited to one, and you add it under Customize → Connectors (support.claude.com, “Getting started with custom connectors using remote MCP”, checked 15 September 2026).
Cost is where the sources disagree. Our own dashboard says “Adding your own connector is a paid feature in both Claude and ChatGPT.” Anthropic’s article reads more generously for a single connector, and OpenAI’s consumer help pages turned us away with a 403 error on 15 September 2026, so check your assistant’s help pages for custom connectors before you plan around a price. The Orbit side is simpler, because there’s no add-on app to buy and no separate charge. On Shopify, WooCommerce or any other platform, ask whether it publishes a connector, and if it does, what that connector can change and how you revoke it.
Do you need an outside assistant at all?
No. Orbit’s own assistant, Nebula, sits inside the dashboard on every plan. Give it a sentence and it adds a product. It also writes product descriptions and blog posts, builds page sections, restyles the shop, runs reports and drafts discounts, and a data question comes back as a chart and a CSV. It can run multi-step plans too, which you approve before they execute. Nebula runs on Claude Sonnet 5.5 by default, with Claude Opus 5.5 selectable when a job deserves more care and images from Google’s Gemini image model.
Its published allowance is roughly 750 product descriptions a month on Solo, 1,500 on Team, 3,750 on Pro and 18,000 on Enterprise, and what each plan costs is on our pricing page. Nebula needs no setup. ChatGPT or Claude, on the other hand, can hold your shop alongside everything else you’ve connected to them, and you can use both.
What does it replace, and what stays with you?
It replaces the tabs you click and the mechanical half of what a virtual assistant does. The judgement stays with you.
Say you have 40 products to add from a supplier spreadsheet, each with a title, description, price, stock figure and two images. We’ll assume 6 minutes a product by hand and £18 an hour for a UK VA; swap in your own figures if they differ. By hand that’s 240 minutes, or 4 hours, which comes to £72. Ask an assistant with safe-write permission to create all 40 as drafts and the job shrinks to one instruction plus roughly 25 minutes of you reading the output and fixing what it got wrong. Call that £7.50 of your own time at the same rate. That’s £64.50 saved on one job, but the 25 minutes didn’t disappear. They’re the check, and the check is the part people skip.
It won’t decide which 40 products are worth stocking, or whether 15% off is the right discount. That was never the tab-clicking part of the job.
Where it earns its keep fastest is the catalogue work behind selling on more than one channel, such as rewriting titles to suit shopping feeds or filling the gaps that stop a product feed validating. Getting enough attributes in to qualify for free Google product listings is more of the same. It’s hours of typing with no thinking in it, which is what you want to hand over. Our guide to multichannel selling in the UK covers the rest.
What can go wrong?
Two things, mainly: it can be confidently wrong, and it can be steered by what it reads. Now and then it will invent a product attribute or miscount a total, and it can misread which spreadsheet column held the price. Read the output on anything that matters, and start any new kind of job with a batch of ten rather than a thousand.
The second risk is the one people underestimate. Our dashboard puts it in two sentences: “An assistant acts on what it reads. Product descriptions, order notes and customer messages are all text that can steer it, so give change permissions only to assistants you trust.” If a stranger can type text into your shop, a stranger can put instructions in front of your assistant. Anthropic makes the same point from the other direction, warning that malicious servers “may include hidden instructions that try to make Claude perform unintended actions” (support.claude.com, checked 15 September 2026). Against this, read permission is close to harmless and live write is not.
Before any big bulk job, export your products, orders and customers as CSV, which Orbit lets you do at any time. It costs nothing and occasionally saves a weekend.
Under our AI assistant data terms you stay the data controller, Orbit is your processor and the assistant provider becomes a further processor. Two lines are worth reading closely. On marketplaces: “You must not use an AI assistant to export, copy, keep or build reports and dashboards from marketplace order data or marketplace buyer data, unless that marketplace allows it.” On revocation: “You can revoke this connection at any time from your dashboard. Revoking stops all future access straight away. It cannot pull back anything the assistant has already received.” The terms also tell you to ask your provider two things before you connect: whether they keep your prompts (and for how long), and whether they use that data to train their models.
Questions people ask about running a shop from ChatGPT or Claude
Can I really run my online shop from ChatGPT or Claude?
You can run the admin side of it. Connect your shop and you can ask questions about sales and stock, create and edit products, build pages, draft discounts and run reports by typing instead of clicking. You still approve anything that moves money or goes out to customers, and you still decide what to sell and at what price.
What is MCP in plain English?
MCP is an open standard that lets an AI application connect to an outside system, described on its own site as “an open-source standard for connecting AI applications to external systems” (modelcontextprotocol.io, checked 15 September 2026). Because it is a standard rather than a private integration, one connection published by your shop platform works with any assistant that supports it.
Is it safe to let an AI assistant change prices or refund orders?
Price changes are fine if you ask for a preview and approve the list first. Refunds we would not automate at all. An assistant acts on the text it reads, including text strangers can write into your shop, so anything irreversible should need a human yes. Orbit splits permissions into read, safe write and live write precisely so you can draw that line.
Does connecting my shop to ChatGPT or Claude cost extra?
Not on the Orbit side: there’s no add-on app to buy and no separate charge. On the assistant side, our dashboard says “adding your own connector is a paid feature in both Claude and ChatGPT”, though Anthropic’s support pages on 15 September 2026 listed custom connectors on Free through Enterprise plans, with free users limited to one. Check your assistant’s help pages, as these rules change often.
Does this replace my virtual assistant?
Partly. It can take over most of the typing: bulk product creation, description rewrites, feed clean-ups and report pulling. Deciding what to stock and answering an upset customer stay with you. Budget time for checking its work, too: on a 40-product import, expect roughly 25 minutes of review.
If you’re weighing up Orbit, run a small version of the worked example before you decide. Ask your assistant for ten draft products from a real supplier spreadsheet, with safe write and nothing more, and count how many you have to fix. That count will tell you more than this page can. The 14-day free trial, no card required, gives you an Orbit shop to run it in: start free.